Model Based System Development/Research Themes/Probabilistic Graphical Models

Probabilistic Graphical Models

Model-based reasoning is an area in artificial intelligence where there is an emphasis on symbolic representation of models of systems and symbolic reasoning with the models, in order to solve problems. Model-based representation and reasoning methods have the virtue that they are understandable, as explicit models with a clear semantics are being developed, and they allow incorporation of quantitative detail when available. A typical example of such models are the causal models used in abductive diagnosis, where a diagnosis of failure of a system is determined by abductive reasoning with the causal knowledge. Many other model-based AI techniques exist, and problems being studied and tacked by these techniques varies from diagnosis to prediction and action planning. The emphasis of our research in model-based reasoning is on combining qualitative and quantitative methods. The motivation underlying this research derives from the fact that in many situations detailed numerical information about real-life processes is difficult, and sometimes impossible, to obtain, whereas it is much easier to acquire information about how variables influence each other qualitatively. Probabilistic graphical models appear to offer a generic framework that allows integrating qualitative and quantitative knowledge, and this framework is taken as the starting point of most of our research. However, logic is also used as a language for modeling, and in some of the research we either draw inspiration from incorporating ideas from logical reasoning to develop probabilistic analogues, or to merge the two languages in one way or other.

A typical example of drawing inspiration from logical concepts for designing new probabilistic methods is our work on the concept of conflict-based diagnosis, that was inspired by existing work on consistency-based reasoning by Johan de Kleer at Xerox at Paolo Alto, where logical consistency checking was replaced by a probabilistic match measure, called the conflict measure. This allowed us to develop a theory similar to conflict-based diagnosis, yet more flexible. Work on integrating logic and probabilistic graphical models was done in exploring the use of Boolean functions in obtaining probability tables defined in terms of causal independence. In addition, we started work on probabilistic logic with the development of a logic that generalizes chain graphs.